Orchestra-Bench

Orchestra-Bench is an English multimodal dataset for high-level cooperative planning among three robots. Each sample provides three views from distinct positions in one scene, a broad user request, and one coordinated text subtask for each robot.

Dataset Overview

• 12,000 samples across 24 scene classes, with 500 samples per class
• 36,000 image references backed by 3,334 unique extracted frames
• Three distinct ground views per sample: ground_1, ground_2, and ground_3
• 17 outdoor classes (8,500 samples) and 7 indoor classes (3,500 samples)
• English user tasks and high-level cooperative robot subtasks
• 14 outdoor source scenes from EgoSchema, contributing 7,000 samples
• Real video frames only; no synthetic or perspective-warped BEV images

Environment	Scene class	Samples
Outdoor	Park path	500
Outdoor	Park junction	500
Outdoor	Building approach	500
Outdoor	EgoSchema golf course 01	500
Outdoor	EgoSchema golf course 02	500
Outdoor	EgoSchema tennis court	500
Outdoor	EgoSchema lawn equipment yard	500
Outdoor	EgoSchema garden work area	500
Outdoor	EgoSchema night road	500
Outdoor	EgoSchema sports field	500
Outdoor	EgoSchema residential road	500
Outdoor	EgoSchema public plaza	500
Outdoor	EgoSchema garden path	500
Outdoor	EgoSchema sidewalk	500
Outdoor	EgoSchema urban walkway	500
Outdoor	EgoSchema grassy field	500
Outdoor	EgoSchema dirt field	500
Indoor	Office and laboratory	500
Indoor	Retail and dining	500
Indoor	Bedroom	500
Indoor	Kitchen	500
Indoor	Living room	500
Indoor	Bathroom	500
Indoor	Corridor	500

Record Format

The main file is manifest.jsonl. Each line has this structure:

{
  "id": "egoschema_dirt_field_0000",
  "scene_class": "egoschema_dirt_field",
  "environment": "outdoor",
  "scene_description": "Outdoor dirt field with uneven ground and sparse vegetation",
  "source_scene": "egoschema/eed1a49f-ba2e-4b83-8817-d8b5d77a3b42",
  "collaboration_pattern": "forward scouting and follow-up",
  "views": [
    {
      "robot_id": "ground_1",
      "view_type": "ground",
      "image": "images/egoschema_dirt_field/egoschema_dirt_field_0000/ground_1.jpg",
      "source_video": "egoschema_eed1a49f-ba2e-4b83-8817-d8b5d77a3b42",
      "timestamp_seconds": 22.0
    }
  ],
  "user_task": "We need to clear a path through the uneven dirt field to reach the target zone ahead.",
  "subtasks": {
    "ground_1": "Move forward to inspect the left side of the path for obstacles or unstable ground.",
    "ground_2": "Approach from behind and check the dirt for soft spots.",
    "ground_3": "Advance while clearing debris and marking the route for the others."
  }
}

Every robot subtask includes an explicit movement action. Plans cover navigation, observation, route checking, blind-spot coverage, reporting, and mutual guidance at a high level rather than low-level controls or trajectories.

Image paths in views are relative to the dataset repository root. Multiple records can reuse a source frame, but the three-view combination within every record is distinct.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "BAAI/Orchestra-Bench",
    data_files="manifest.jsonl",
    split="train",
)
print(dataset[0])

This dataset is the result of joint research conducted by Hao Tang's team at the School of Computer Science, Peking University, and the Beijing Academy of Artificial Intelligence (BAAI).